Assignment 3

Step 1 - Prepare the Data

Resize all of the images before the training to save time

exmpl.png

Step 2 - Create an initial model

epochs_model_1.png

loss_model_1.png

It can be seen that the training loss decrese at every epoch, but the validation loss goes up and down, could be a problem with the learning rate. overall the validation loss was lower then train loss at all times.

psnr_model1.png

The PSNR increases at every epoch, so it looks like the model did learn

imgs_model_1.png

Step 3 - Add block to the initial model

epochs_model_@.png

loss_model_2.png

The validation loss decreased at most of the time, looked better than the first model. The train loss is decreasing every epoch.

psnr_model_2.png

The PSNR was increasing at every epoch and got to same results as the previous model

mid_img_model_2.png

large_img_model_2.png

After the first two epochs the images got weird colors, but as the model continue the colors got better

Step 4 - Add residual block

epochs_model_3.png

loss_model_3.png

In this model it was the first time that the validation loss was decreasing at every epoch. The train loss was going down as it did in previous models.

psnr_model_3.png

The PSNR was increasing at every epoch and got better results then the two previous models.

mid_img_model_3.png

large_img_model_3.png

It can be seen that this model got better result than the previous two, the images are less blury.

Step 5 - Replace residual block with dilated (Atrous) convolutional block

epochs_model_4.png

loss_model_4.png

The validation and train loss were both decreasing which is great :)

psnr_model_4.png

The PSNR was increasing every epoch but the results were better at the previous model.

mid_img_model_4.png

large_img_model_4.png

For some reason the middle image was grey at first but with each epoch the colors got better. The large image started very blurry with weird colors but with every epoch got better.

Step 6 - Pretrained model

epochs_model_5.png

loss_model_5.png

With the VGG16 net the train loss dropped very fast at start and got stabled. The validation loss starts low and dropped a little.

psnr_model_5.png

This model got the higher PSNR result

mis_img_model_5.png

large_img_model_5.png

This model was less blurry than the other models, and the colors were correct through the whole epochs.

Step 8 - Improve one of the above models

I decided to combine two models, the pretrained vgg16 and the residual-block model

epochs_model_6.png

loss_model_6.png

Also the validation loss was lower than the train loss but it looks like the loss was increasing most of the time and the train loss is decreasing. So the model could got overfit.

psnr_model_6.png

Still this model got the best PSNR results.

mid_img_model_6.png

large_img_model_6.png

As we seen in the previous model, the images was less blurry then in the others model and the colors were good at all times.

table.png